Multiscale Feature Fusion by Entropy-Augmented KCFP for Rolling Bearing Fault Diagnosis

Deep learning models for bearing fault diagnosis often struggle to preserve discriminative features under variable operating conditions and unable to effectively capture multi-scale signal patterns. To overcome these challenges, Entropy-Augmented Kronecker Convolutional Feature Pyramid (EKCFP) was proposed, a novel framework with three core contributions. First, we developed a permutation entropy fusion, a method that directly combines signal complexity analysis with raw vibration inputs via dedicated channels, enabling simultaneous learning of temporal dynamics and information-theoretic properties. Second, a Frequency-Spatial Convolution layer (1DFSC) jointly used dilated convolution and channel attention to extract localized spatial relationships and feature fusion. Third, CNN-based Kronecker Convolution Feature Pyramid (KCFP) integrated batch normalization, adaptive pooling, and stochastic dropout within a compact architecture, achieving greater parameter efficiency. Rectified Linear Unit (ReLU) was used to improve training in order to increase network representation and efficiency. Tests on the HFZZ dataset (Shandong University Intelligent Sensors Center) and Paderborn University bearing data across several fault types show EKCFP achieves 99.94% mean accuracy—exceeding existing multi-scale feature extraction techniques. Ablation experiments highlighted the entropy channel and attention mechanism significance, boosting diagnostic consistency under variable operating conditions. These findings established a novel methodological framework for integrating entropy-driven feature learning with spatial-frequency analysis, advancing the development of explainable deep architectures for mechanical fault diagnostics.